Aim 1 (Experimental)

The first aim of my final project is to build and simulate a genome-scale metabolic and circuit-level ODE model of the four-module ÌṢỌ architecture by utilising Tellurium/libroadrunner for time-course simulation, SALib for global sensitivity analysis, and NumPy-based Moran process modelling for evolutionary containment stability, generating a Pareto-resolved fitness-efficacy landscape and a ranked parameter influence analysis as the primary computational output.

Aim 2 (Developmental)

Following a successful Aim 1, the top-ranked parameter regimes from the Pareto landscape will guide assembly and transformation of the sense-respond-contain circuit into EcN. The engineered sentinel will be tested in co-culture assays against Salmonella Typhimurium and E. coli O157:H7, validating both the tetrathionate-sensing threshold and MccH47-mediated kill kinetics experimentally. Discrepancies between model predictions and wet-lab data will feed back into model refinement.

Aim 3 (Visionary)

The long-term goal is a rugged, orally delivered live biotherapeutic that operates autonomously in the gut, activates only in the presence of pathogen-associated tetrathionate, kills narrow-spectrum without collateral microbiome disruption, and cannot persist outside the host. If the fitness-governability framework holds, ÌṢỌ becomes a design methodology applicable beyond this specific pathogen set, with direct relevance for AMR management, inflammatory bowel disease, and cancer immunotherapy in low-resource clinical settings.

Literature context

Palmer et al. (2017, ACS Infectious Diseases) demonstrated that EcN can be engineered to sense gut-luminal tetrathionate via the TtrS/TtrR two-component system and produce Microcin H47 in response, achieving measurable Salmonella inhibition in a mouse colonisation model. Critically for ÌṢỌ, this paper provides experimentally validated, ODE-parameterisable values for sensor activation kinetics, MccH47 production rates, and pathogen kill constants, making it the direct quantitative predecessor to this project rather than simply a conceptual reference.

Stritzker et al. (2007, International Journal of Medical Microbiology) characterised deltaDAPA auxotrophy in EcN in detail, reporting an escape frequency of approximately 10^-8 per generation under DAP-free conditions. That specific number is what makes the containment module computationally tractable: escape probability can be directly parameterised in the Moran process model rather than estimated from first principles.

What is novel

What ÌṢỌ does that neither paper does is treat fitness cost as a first-class design variable rather than a post-hoc observation. Every published EcN engineering study acknowledges metabolic burden; none model it explicitly as a design input alongside efficacy. ÌṢỌ builds a Pareto frontier that makes the tradeoff navigable rather than anecdotal. The containment module also moves from binary characterisation (auxotrophy is present or absent) to a dynamic system property, asking how quickly a loss-of-function mutant fixes in a finite population over evolutionary time.

Why this matters

Diarrhoeal disease causes approximately 1.6 million deaths per year globally, with the under-five burden concentrated in West and East Africa. Nigeria alone accounts for a disproportionate share of this mortality. Existing interventions reduce severity but do not prevent recurrence in high-transmission settings, and empirical antibiotic use is accelerating resistance emergence in the pathogens most responsible for paediatric deaths: Salmonella, enterotoxigenic E. coli, and Shigella. A sentinel probiotic that activates conditionally, kills narrow-spectrum, and cannot persist outside the host addresses this without adding to AMR pressure.

Beyond the immediate clinical problem, the fitness-governability framework ÌṢỌ develops has broader implications. Any engineered living therapeutic faces the same core question: will the circuit hold under the evolutionary pressure of a real biological environment? Current regulatory frameworks for live biotherapeutics have no standardised computational tool for answering this before a clinical trial. ÌṢỌ begins building one. Nigerian and broader West African epidemiological data (Egbewale 2022; Gayawan 2024) are used to parameterise disease burden and clinical context from the start, not as a framing afterthought.

Ethical implications

Two principles are directly engaged here: beneficence and justice. A precision antimicrobial that spares the commensal microbiome and cannot persist outside the host is strictly better than empirical broad-spectrum antibiotics for the patient, for the microbiome, and for the resistance landscape. Research that addresses paediatric mortality in West Africa while remaining computationally grounded in West African epidemiology represents a genuine departure from the default of developing interventions for high-income contexts and adapting them downstream.

The risks require honesty. A single deltaDAPA deletion is probably not sufficient for any real-world deployment. The current model assumes a closed population and does not account for horizontal gene transfer of the dapA gene from environmental bacteria. The Moran process also excludes commensal competition dynamics, so estimates of circuit persistence are optimistic. These are known limitations, explicitly scope-bounded to this computational phase. Non-maleficence requires that these caveats travel with any communication of the results. Open-source model release via GitHub (MIT licensed) is a deliberate act toward equitable access to the methodology.

Modelling pipeline

Circuit topology → ODE construction → parameter sweep → Pareto analysis → sensitivity analysis → evolutionary stability

Biosensor signal

Chosen: tetrathionate via TtrS/TtrR two-component system. Pathogen-specific: Salmonella and E. coli O157:H7 produce tetrathionate during gut inflammation via reactive oxygen species. Experimentally validated in EcN (Palmer et al. 2017). Signal is absent under homeostatic conditions, directly minimising leaky expression burden at baseline.

Microcin effector

Chosen: Microcin H47 (MccH47). Naturally produced by EcN. Narrow-spectrum: E. coli, Salmonella, Shigella. Mechanism is ATP synthase inhibition, a well-characterised mode of action enabling direct ODE kill-kinetics parameterisation. Immunity protein MchI is endogenous to the EcN chassis. Palmer 2017 provides benchmarked production and kill-rate values for exactly this design.

Containment

Chosen: deltaDAPA auxotrophy (diaminopimelic acid / DAP). DapA is essential for lysine and peptidoglycan synthesis. DAP is absent from the mammalian gut: no dietary source, no commensal production. Deletion is lethal without exogenous supply. Validated in EcN (Stritzker et al. 2007). Published escape frequency ~10^-8 per generation is directly parameterisable for the containment escape model.

ODE framework

Chosen: Tellurium + libroadrunner (SBML/Antimony). Purpose-built for systems biology ODE modelling. Antimony syntax maps directly onto circuit topology (promoter to mRNA to protein). libroadrunner’s stiff CVODE solver handles fast mRNA turnover and slow protein accumulation dynamics without manual configuration. SBML export makes every model citable and reproducible. SciPy solve_ivp (LSODA flag) runs in parallel for parameter sweeps and Pareto grid computation.

Sensitivity analysis

Primary: PRCC via SALib (Marino et al. 2008). Designed for nonlinear, monotonic systems, exactly what Hill-function gene circuits produce. 500 to 2000 Latin hypercube samples sufficient for 6 to 8 parameters.

Supplementary: Sobol total-order indices. Captures interaction effects (Hill coefficient n and KD interact in the sensor module). 5000 to 10000 samples, tractable on a laptop in minutes.

Evolutionary stability

Chosen: Moran process with fitness-weighted selection. Two competing types: functional circuit (fitness 1 minus delta) and loss-of-function mutant (fitness 1). Fixation probability computed analytically (Nowak 2006), then 1000 stochastic trajectories via numpy.random.choice() with fitness-weighted birth-death events. Directly answers: how long does the circuit remain functional under selection pressure?

Environment setup

python3.11 -m venv ~/.envs/iso-ecn && \
source ~/.envs/iso-ecn/bin/activate && \
pip install --upgrade pip && \
pip install tellurium scipy numpy \
    matplotlib seaborn SALib pandas && \
pip freeze > requirements.txt

Task 1: Environment setup and baseline biosensor model (weeks 1)

Install and configure the full modelling stack: Tellurium, libroadrunner, SALib, NumPy, Matplotlib, Seaborn, SciPy, all pinned in a venv with requirements.txt. Build the biosensor module as a two-ODE Hill-function model encoding the TtrS/TtrR tetrathionate-to-promoter activation pathway. Fit activation threshold KD and Hill coefficient n against Palmer 2017 time-course data.

Expected result: Simulated sensor activation curve matches digitised Palmer 2017 experimental data within 20% across the measured tetrathionate concentration range.

Task 2: Full four-module ODE construction (weeks 2)

Extend the biosensor ODE to include the regulator module (thresholded Hill-function promoter gating effector expression), the effector module (MccH47 production and pathogen kill kinetics), and the containment module (deltaDAPA escape probability). Write all models in Antimony syntax within Tellurium. Export validated models to SBML and commit to GitHub.

Expected result: Stable steady-state solutions for all four modules under both homeostatic and pathogen-present conditions. Leaky expression at baseline should approach zero.

Task 3: Pareto landscape and parameter sweep (weeks 3)

Use SciPy solve_ivp with LSODA flag to sweep burden parameter delta and effector output rate k_M across a 50 x 50 parameter grid. Record steady-state growth rate and pathogen suppression ratio for each grid point. Plot Pareto frontier, colour-coded by regulator variant (linear vs. thresholded).

Expected result: A visible Pareto frontier separating viable design space from over-burdened and under-effective regions. The thresholded regulator variant should dominate the frontier.

Task 4: Global sensitivity analysis (weeks 4)

Run PRCC analysis via SALib using Latin hypercube sampling across 6 to 8 parameters: Hill coefficient n, signal threshold KD, burden delta, MccH47 production rate k_M, pathogen kill rate k_kill, mRNA degradation rate, and protein dilution rate. Generate ranked tornado chart. Run supplementary Sobol total-order index analysis.

Expected result: n and KD rank as the top two PRCC drivers of sensor module output. Sobol indices confirm a significant interaction effect between the two.

Task 5: Evolutionary stability via Moran process (weeks 5)

Implement the Moran process in NumPy. Define two competing cell types (functional circuit, fitness 1 minus delta; loss-of-function mutant, fitness 1). Compute analytical fixation probability from Nowak 2006. Run 1000 stochastic trajectories. Vary delta across the Pareto-viable range; plot fixation probability across three population sizes with analytical solution overlaid.

Expected result: Fixation probability of the loss-of-function mutant increases sharply above delta = 0.1. This is the quantitative argument for why the thresholded regulator module is not optional.

Industry Council companies

What is being validated

The primary validation for this phase is the biosensor module ODE fitted to Palmer 2017 experimental data, and the resulting Pareto landscape generated by the four-module parameter sweep. Together, these demonstrate that the computational framework is grounded in real parameterisation, not arbitrary simulation.

Validation protocol

1. Extract tetrathionate-responsive GFP expression data from Palmer et al. 2017 Figure 3 (digitised using WebPlotDigitizer).
2. Build the TtrS/TtrR two-component ODE in Antimony syntax within Tellurium. Initial parameter estimates: Hill coefficient n = 2.0, activation threshold KD = 50 micromolar.
3. Run time-course simulations across five tetrathionate concentrations matching Palmer 2017 experimental conditions.
4. Fit KD and n using SciPy curve_fit with residual sum of squares minimisation. Report fitted values with 95% confidence intervals.
5. Validate that simulated activation curve matches digitised data within 20% across the concentration range.
6. Extend validated sensor ODE to the full four-module system. Run 50 x 50 parameter sweep. Record steady-state outputs for each grid point.
7. Identify and plot the Pareto frontier. Export 300 dpi PNG and SVG to GitHub.
8. Commit all code, SBML files, and figures to Jonahnki/iso-sentinel-ecn under MIT licence.

Challenges and strategies

The main challenge so far is parameter identifiability in the effector module. The MccH47 kill-rate constant k_kill is not directly reported in Palmer 2017; it is inferred from pathogen viable-count time-courses. The current approach uses a range of literature-sourced bacteriocin kill-rate values (10^{6 to 10}8 cells per micromolar per hour) and propagates this uncertainty explicitly through the sensitivity analysis rather than selecting a single point estimate. This produces confidence intervals on the Pareto frontier that reflect real parametric uncertainty rather than false precision.

A second known limitation is the absence of commensal competition in the current Moran process model. Current fixation probability estimates are therefore optimistic by design. This is flagged explicitly in every figure caption involving evolutionary stability output. The Lotka-Volterra competition extension planned for post-course development is the right fix; it is out of scope here and is stated as such.

Distinction from existing work

The engineered probiotic field asks: can we build a circuit that works?

ÌṢỌ asks: across what design regimes does a circuit remain both functional and governable under the pressures that will actually be present?

• Fitness cost as a design variable — no published EcN paper produces a fitness-efficacy Pareto frontier. All existing work acknowledges burden; none model it as a first-class input.
• Containment as a dynamic system property — the field treats auxotrophy as binary. ÌṢỌ models escape probability over evolutionary time via the Moran process.
• Disease context — the dominant literature targets IBD and colorectal cancer. ÌṢỌ is framed around acute paediatric diarrhoeal disease in West African clinical settings; this changes which signals, effectors, and ecological assumptions are relevant.
• Model-first methodology — existing EcN engineering papers build constructs first and measure them. ÌṢỌ maps the computational design landscape before any construct is built.
• Geographic grounding — clinical inspiration and epidemiological parameters drawn from Nigerian data (Egbewale 2022; Gayawan 2024). African-origin disease burden as a scientific foundation, not a framing afterthought.

What ÌṢỌ builds on

• Palmer et al. 2017 — direct experimental predecessor; tetrathionate/MccH47/EcN parameter source
• Weibel et al. 2024 — modular architecture precedent; ÌṢỌ extends with containment module and ODE-level analysis
• Ba et al. 2023 — EcN toolbox that any future wet-lab build would draw on
• Stritzker et al. 2007 — deltaDAPA escape frequency; containment parameterisation source
• Marino et al. 2008 — PRCC methodology; sensitivity analysis canonical citation

Repository

Jonahnki/iso-sentinel-ecn

All model code, SBML files, and figures. MIT licensed. CITATION.cff included.

1. Palmer, J. D., Piattelli, E., McCormick, B. A., Silby, M. W., Brigham, C. J., and Bucci, V. (2017). Engineered probiotic for the inhibition of Salmonella via tetrathionate-induced production of microcin H47. ACS Infectious Diseases, 4(1), 39-45. https://doi.org/10.1021/acsinfecdis.7b00114

2. Weibel, N., Curcio, M., Schreiber, A., et al. (2024). Engineering a novel probiotic toolkit in Escherichia coli Nissle 1917 for sensing and mitigating gut inflammatory diseases. ACS Synthetic Biology, 13(8), 2376-2390. https://doi.org/10.1021/acssynbio.4c00036

3. Lynch, J. P., Goers, L., and Lesser, C. F. (2022). Emerging strategies for engineering Escherichia coli Nissle 1917-based therapeutics. Trends in Pharmacological Sciences, 43(9). https://doi.org/10.1016/j.tips.2022.02.002

4. Ba, F., Zhang, Y., Ji, X., Liu, W.-Q., Ling, S., and Li, J. (2023). Expanding the toolbox of probiotic Escherichia coli Nissle 1917 for synthetic biology. bioRxiv. https://doi.org/10.1101/2023.06.05.543671

5. Stritzker, J., Weibel, S., Hill, P. J., Oelschlaeger, T. A., Goebel, W., and Szalay, A. A. (2007). Tumor-specific colonisation, tissue distribution, and gene induction by probiotic E. coli Nissle 1917 in live mice. International Journal of Medical Microbiology, 297(3), 151-162.

6. Marino, S., Hogue, I. B., Ray, C. J., and Kirschner, D. E. (2008). A methodology for performing global uncertainty and sensitivity analysis in systems biology. Journal of Theoretical Biology, 254(1), 178-196.

7. Nowak, M. A. (2006). Evolutionary Dynamics: Exploring the Equations of Life. Harvard University Press.

8. Moran, P. A. P. (1958). Random processes in genetics. Mathematical Proceedings of the Cambridge Philosophical Society, 54(1), 60-71.

9. Egbewale, B. E., Karlsson, O., and Sudfeld, C. R. (2022). Childhood diarrhea prevalence and uptake of oral rehydration solution and zinc treatment in Nigeria. Children, 9(11), 1722.

10. Gayawan, E., Cameron, E., Okitika, T., Egbon, O. A., and Gething, P. (2024). A situational assessment of treatments received for childhood diarrhoea in the Federal Republic of Nigeria. PLOS ONE, 19(5), e0303963.

11. World Health Organization. (2024). Diarrhoeal disease. https://www.who.int/news-room/fact-sheets/detail/diarrhoeal-disease